I ran into something interesting while using AI.
You spend weeks working with an AI on a project. It learns your preferences, understands your approach, remembers past decisions, and gets better at helping you.
Then, a few weeks later, you return to that same project with a new problem.
The AI remembers what you were building. It remembers the decisions you made. But it also carries forward assumptions that no longer apply. Instead of helping you move forward, its memory starts getting in the way.
And that got me thinking.
Does giving AI more memory actually make it smarter, or does it just give it more context to make assumptions with?
Forgetting too much makes us repeat ourselves. Remembering too much can make AI cling to outdated information.
Where’s the balance?
Would you rather have an AI that remembers almost everything about your work, or one that remembers only what’s relevant to the problem at hand?
Curious to hear how this plays out in your actual workflows.
I don’t think it makes it smarter - you are pruning its search space when sampling, you are changing its prior so next time it does something for you it has an inherent bias and knows what NOT to explore. I.e. you are creating personalisation and providing it a different context apriori.
Apologies if you’ve seen my related posts already! But a good RAG data store, a good combination of pipeline and agentic retrieval, and carefully-crafted project instructions get you light years beyond built-in ‘memory and cross-chat looks’ on any platform I’ve so far used.
@platypus I like this distinction.
Memory might not make the model smarter in itself, but it changes the context in which it reasons and the directions it is more likely to explore.
The interesting question is whether better personalization eventually looks like intelligence from the user’s perspective
I totally agree here.
The interesting part is that better results might come less from making the model remember more and more from giving it the right information at the right time. But again has it’s on limitations.